Scholarlink

<p>Timely and accurate neuroimaging interpretation is fundamental to management strategies for traumatic brain injuries (TBI), with early diagnostic information having a critical role in influencing outcomes in patients. Most previous automatic systems, nonetheless, utilize static imaging and do not include temporal dynamics and contextual subtlety essential to neurological development. In this work, we propose a time-conscious multimodal deep learning system for automatic radiology reporting in TBI diagnosis. Our framework integrates an Augmented Convolutional Bi-directional Feature Pyramid Network (AC-BiFPN) for multi-scale image representation with a Transformer-based decoder to accommodate longitudinal imaging sequences and structured clinical metadata (e.g., age, sex, neurological history). By simulating temporal sequences of CT/MRI, our model is able to identify subtle disease development patterns such as lesion regression or exacerbation. Tested on the RSNA Intracranial Hemorrhage Detection dataset, our method outperforms robust baselines in natural language generation and on clinical measures including CheXpert-derived validation. Qualitative results also illustrate that the model can generate clinically coherent narratives concurring with the temporal course of neurological abnormalities. This research closes the gap between visual intelligence and brain informatics by presenting a scalable and cognitively-inspired method for real-time neurodiagnostic reporting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Temporal multimodal transformer for automated radiology reporting in traumatic brain injuries

  • 期刊论文

  • 发表日期：

    2025年8月2日

查看全文
  • Riadh Bouslimi,
  • Mariem Medini,
  • Maissa Ben Fradj

摘要

Timely and accurate neuroimaging interpretation is fundamental to management strategies for traumatic brain injuries (TBI), with early diagnostic information having a critical role in influencing outcomes in patients. Most previous automatic systems, nonetheless, utilize static imaging and do not include temporal dynamics and contextual subtlety essential to neurological development. In this work, we propose a time-conscious multimodal deep learning system for automatic radiology reporting in TBI diagnosis. Our framework integrates an Augmented Convolutional Bi-directional Feature Pyramid Network (AC-BiFPN) for multi-scale image representation with a Transformer-based decoder to accommodate longitudinal imaging sequences and structured clinical metadata (e.g., age, sex, neurological history). By simulating temporal sequences of CT/MRI, our model is able to identify subtle disease development patterns such as lesion regression or exacerbation. Tested on the RSNA Intracranial Hemorrhage Detection dataset, our method outperforms robust baselines in natural language generation and on clinical measures including CheXpert-derived validation. Qualitative results also illustrate that the model can generate clinically coherent narratives concurring with the temporal course of neurological abnormalities. This research closes the gap between visual intelligence and brain informatics by presenting a scalable and cognitively-inspired method for real-time neurodiagnostic reporting.